Transitory cross entropy for model training on unbalanced datasets
2022pp. 17–17
Abstract
The proposed transitory cross entropy loss function performs a weighted average of the cross entropy using both the truth labels and the predicted labels; this is a variation of the weighted cross entropy loss function that performs a weighted average using just the truth labels. We tested the transitory cross entropy loss function by training ICNet on the CityScapes dataset and saw an increase in the mean-intersection-over-union relative to the model trained using the standard weighted cross entropy loss function. We further propose modifying the weights based on dynamic performance metrics rather than just static distribution metrics.